Hidden Markov Model (HMM) based Speech Synthesis for Urdu Language Omer Nawaz Dr. Tania Habib Centre for Language Engineering, Al- Khawarizmi Institute of Compute Science, UET, Lahore, Pakistan. [email protected]Computer Science and Engineering Department UET, Lahore, Pakistan. [email protected]Abstract This paper describes the development of HMM based speech synthesizer for Urdu language using the HTS- toolkit. It describes the modifications needed to original HTS-Demo-scripts to port them, for Urdu language, which are currently available for English, Japanese and Portuguese. That includes the generation of the full- context style labels and the creation of the Question file for Urdu phone set. For that the development and structure of utilities are discussed. Plus a list of 200 high frequency Urdu words are selected using the greedy search algorithm. Finally the evaluation of these synthesized words is conducted using naturalness and intelligibility scores. Keywords— Speech Synthesis, Hidden Markov Models (HMMs), Urdu Language, Perceptual Testing 1. Introduction A text-to-speech (TTS) synthesis system for a particular language is a framework to convert any given text into its equivalent spoken waveform representation. Currently the most frequently employed TTS is the Unit Selection Synthesis [1-3]. However being the best TTS to date it has some limitations. Like the synthesized speech resembles the prosody/style of recording with the training database. If we want to synthesize speech with various voice characteristics then we need to increase the training data that cover all that variations. However recording that much data is not feasible [4]. With the improvements in Hidden Markov Models (HMM) techniques, the HMM based speech synthesizers are becoming popular [5]. In these systems the statistical models are trained based on source filter model from the training corpus. The main advantage of parametric approach [6] is that original waveforms are not required to be stored for synthesis purposes. As an 1 Footprint refers to the amount of disk space required by an application. effect the foot-print 1 is very small (approximately 2- MB 2 ), compared to unit selection approach. The HMM-based speech synthesis framework has been applied to a number of languages that include English [7], Chinese [8], Arabic [9], Punjabi [10], Croatian [11] and Urdu [12] as well. In this work, we present the development and evaluation of Speech Synthesizer for Urdu language. The main contributions of the paper are inclusion of prosodic information in the training process and development of question set considering the linguistic features relevant to Urdu language. Figure 1 depicts the outline of parametric speech synthesis with HMMs. The training part consists of extracting the feature vectors of the training corpus as mel-cepstral coefficients [13] and excitation parameters, followed by model training. Whereas synthesis part is the reverse process of speech recognition. First the text is converted to context dependent sequence of phones obtained as a part of Natural Language processing (NLP) [14]. Then the excitation and spectral parameters are obtained through a set of trained HMM models using parameter generation algorithm [15]. Finally the waveform is generated using the obtained spectral and excitation features and providing them to the mel-log spectrum approximation filter (MLSA) [16]. 2 This refers to the voice size produced by the HTS- English-Demo Scripts. Excitation SPEECH DATABASE Excitation Parameter extraction Spectral Parameter Extraction Excitation generation Synthesis filter TEXT Text analysis SYNTHESIZED SPEECH Parameter generation from HMMs Context-dependent HMMs & state duration models Labels Excitation parameters Spectral parameters Speech signal Training part Synthesis part Training HMMs Spectral parameters Excitation parameters Labels Figure 1. Overview of Parametric speech synthesis with HMMs ( [7], pp. 227)
8
Embed
Hidden Markov Model (HMM) based Speech Synthesis for Urdu …cle.org.pk/Publication/papers/2014/Hidden Markov Model... · 2014-10-16 · Hidden Markov Model (HMM) based Speech Synthesis
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Hidden Markov Model (HMM) based Speech Synthesis for Urdu Language
Omer Nawaz Dr. Tania HabibCentre for Language Engineering, Al-
2.4.5. Tree traversal for model selection. For example
if we want to synthesize the word P A K I S T A N,
then for each phoneme a separate tree will be used to
trace the appropriate model. In this example we may
have two different models for the ‘A’ phoneme, because
it is occurring twice and have different left and right
contexts. For the case of ‘P A K’, first it checks whether
left context is bilabial or not (L=Bilabial?). Then on its
outcome it proceeds to the next node, and finally reaches
the leaf node from where suitable model is selected as
shown in Figure 5.
Figure 5. Tree traversal for model selection
3. Evaluation and results
The main goal of any Text to Speech System is to
generate a voice which resembles closely to a human
voice. So for the assessment of a speech synthesizer, a
human listener should carry out the testing.
To test the system comprehensively, there are a number
of tests, which include Diagnostic Rhyme Test (DRT)
[28] and Modified Diagnostic Rhyme Test (M-DRT)
[29] that evaluates the system on the phoneme level.
For our system we only focused on the
naturalness and intelligibility of high frequency words.
As Consonant-Vowel-Consonant (CVC) [30] or DRT
was not appropriate because the number of possible
correct words fitting exactly the CVC format was
scarce. Moreover, we did not have the phoneme
coverage balanced for the 30-minutes of the speech data.
The phoneme coverage graph is shown in Figure 6.
Consequently a set of 200 high frequency words were
selected for the testing purpose.
3.1. Methodology
To perform evaluation of the underlying system a
list of 200 high frequency words of Urdu language were
selected using the greedy search algorithm [31]
developed at Center for Language Engineering, KICS.
3.2. Experiment
For the assessment of speech quality synthesized
by the statistical models (HMMs). The Mean-Opinion-
Score (MOS) was considered for the naturalness and
intelligibility measure. There were a total of 4 listeners
who carried out the evaluation. Among the participants,
three were linguists (expert listeners) and one was
technical (naive listener).
For our system the naturalness and intelligibility are
interpreted as:
Naturalness: How close it seems to be produced by
a human?
Intelligibility: How much conveniently the word
was recognized?
The MOS-scale varies from 1 to 5, where 1 represents
the lowest score and 5 the highest. The experimental
results of four listeners are listed in Table 2:
Table 2. Mean score for Intelligibility and Naturalness
Subject Type MOS
Naturalness
MOS
Intelligibility
Technical 1 3.23 3.65
Linguistic 1 2.82 3.66
Linguistic 2 2.86 3.58
Linguistic 3 3.48 3.52
The testing reveals that most of the words were
intelligible but not natural. The reason behind un-natural
voice can be regarded due to the kind of training data.
In training, words were available as a carrier sentence,
and none of the training utterance consisted of a single
word. We know that if a word is spoken explicitly
without any carrier sentence then it is little bit longer
and clearer, whereas in carrier some of the phonemes are
shorter or are completely ignored.
4. Analysis and Discussion
The analysis show that on average 92.5% words
were correctly identified, irrespective they sounded less
natural or intelligible. On the other hand, there were also
a few cases where the listener was unable to identify the
correct word. These are listed in Table 3.
L = Bilabial?
C=stop? R = Stop?
L=Voiced?
L=Vowel?
R=Voiced?
yes no
yes
no
no yes
no yes
Leaf Leaf
Sub-Tree
|P |A |K| L = Bilabial yes
R = Stop yes L = Voiced no L = Vowel no R = Voiced no
Figure 6. Phoneme Coverage Graph4
4.1. Phoneme coverage of training data
There were total of 66 different phonemes present
in the phone set defined for training HMMs Models,
having total frequency of 17793 (30-minute data). So for
completely phonetically balanced system we should
have at least 270 (1.51 % coverage) examples per
phoneme. Whereas in our case vowels had a very high
frequency (A = 1810, A_A = 1646), and some of the
consonants were completely ignored (J_H, L_H, M_H,
N_G_H, R_H, Y, Z_Z). The phoneme coverage can be
visualized in Figure 6.
A list of words that were not correctly identified are
listed in Table 3.The first column contains the word in
Nastalique style. Second represents the actual
pronunciation that should have been synthesized, in
CISAMPA format. Whereas third column represents the
word interpreted by the listener. The bold letters
highlights the phones which have disagreements, while
gray letters indicates that they were missing in the
synthesized utterance. Finally the last column represents
the coverage of the correct phoneme that was wrongly
produced, in the training corpus.
Table 3. Words with errors
Nastalique
Style
CISAMPA
(Correct)
Listened
(Incorrect)
Coverage
(%)
رطفT_DARAF T_DALAF 5.92
اگGA_A D_DA_A 1.35
ولعممMAYLU_UM MAT_DLU_UM 0.00
ےھتT_D_HA_Y T_SA_Y 0.66
رزیRAZI_I RAD_DI_I 0.88
وہیتHO_OT_DI_I HO_OT_DI_I 4.68
ویکہکنKIU_U_NKA_Y T_SU_NKA_Y 0.15
4 Only those phonemes are shown whose occurrence
counts are more than 50
قحHAQ HABS 0.46
دعبBAYD_D BAYD_D 0.00
ایخلXAJA_AL FIJA_AL 0.50
5. Conclusion and Future Work
A reasonably good quality5 HMM based speech
synthesizer for Urdu language has been developed. The
utilities developed were unique as they converted hand-
labeled TextGrid files directly to HTS-label format,
without using any of the automatic data tagging
software (like Sphinx [32]). The Question file was
generated for the Urdu phone set, keeping in account the
articulatory features of language. Finally the testing of
the synthesized quality was carried out by using the
Mean-Opinion-Score (MOS) for naturalness and
intelligibility.
In future work we are planning to build the system
incrementally with new database which comprises of
approximately 10-hours of speech and is being recorded
by a professional speaker
Acknowledgments
This work has been conducted through the project,
Enabling Information Access for Mobile based Urdu
Dialogue Systems and Screen Readers supported
through a research grant from ICTRnD Fund, Pakistan.
References [1] A. W. Black and P. Taylor, "CHATR: a generic speech synthesis system," in proc. of the 15th conference on Computational linguistics, Stroudsburg, PA, USA, 1994.
5 Some synthesized utterances can be accessed at:
http://www.cle.org.pk/tts/sample
0
500
1000
1500
2000
A
A_
A
A_
Y R K I_I H N M S
T_
D L B I
O_O
D_D U P G
D_Z J
A_
E V
A_
Y_N O
T_
S
U_U Z
O_O
_N
T_
D_H
R_R
K_
H
S_
H F X Q T
T_
S_H
B_
H E
A_
A_N
A_
E_N
I_I_
N
[2] R. E. Donovan and P. C. Woodland, "Automatic speech synthesiser parameter estimation using HMMs," in proc. of ICASSP-95, Detroit, Michigan, May, 1995. [3] A. J. Hunt and A. W. Black, "Unit selection in a concatenative speech synthesis system using a large speech database," in proc. of ICASSP-96, IEEE International Conference, Atlanta, Georgia, May, 1996. [4] A. W. Black, "Unit selection and emotional speech," in proc. of INTERSPEECH, Geneva, September, 2003. [5] Z. Heiga, T. Tomoki, M. Nakamura and K. Tokuda, "Details of the Nitech HMM-based speech synthesis system for the Blizzard Challenge 2005," IEICE transactions on information and systems, vol. 90, no. 1, pp. 325--333, 2007. [6] H. Zen, K. Tokuda and A. W. Black, "Statistical parametric speech synthesis," Speech Communication, vol. 51, no. 11, pp. 1039-1064, 2009. [7] K. Tokuda, H. Zen and A. W. Black, "An HMM-based speech synthesis system applied to English," in Speech Synthesis, 2002. Proceedings of 2002 IEEE Workshop, Santa Monica, California, IEEE, September, 2002, pp. 227-230. [8] Y. Qian, F. Soong, Y. Chen and M. Chu, "An HMM-based Mandarin Chinese text-to-speech system," in Chinese Spoken Language Processing, vol. 4274, Springer Berlin Heidelberg, Singapore, December, 2006, pp. 223-232. [9] O. Abdel-Hamid, S. M. Abdou and M. Rashwan, "Improving Arabic HMM based speech synthesis quality," in proc. of INTERSPEECH, Pittsburgh, Pennsylvania, USA, September 17-21, 2006. [10] D. Bansal, A. Goel and K. Jindal, "Punjabe speech synthesis using HTK," International Journal of Information Sciences & Techniques, vol. 2, no. 4, July, 2012, pp. 57-69. [11] I. Ipsic and S. Martincic-Ipsic, "Croatian HMM-based speech synthesis," CIT. Journal of computing and information technology, vol. 14, no. 4, December, 2006, pp. 307-313. [12] Z. Ahmed and J. P. Cabral, "HMM BASED SPEECH SYNTHESISER FOR THE URDU LANGUAGE," in 4th International Workshop On Spoken Language Technologies For Under-resourced Languages, St. Petersburg, Russia, 2014. [13] T. Fukada, K. Tokuda, T. Kobayashi and S. Imai, "An adaptive algorithm for mel-cepstral analysis of speech," in proc. of ICASSP-92., IEEE International Conference, San Francisco, California, March, 1992. [14] H. Kabir, S. R. Shahid, A. M. Saleem and S. Hussain, "Natural Language Processing for Urdu TTS System," in Multi Topic Conference, 2002. Abstracts. INMIC 2002. International, IEEE, 2002, pp. 58-58. [15] K. Tokuda, T. Yoshimura, T. Masuko, T. Kobayashi and T. Kitamura, "Speech parameter generation algorithms for HMM-based speech synthesis," in proc. of ICASSP'00, IEEE International Conference, Istanbul, Turkey, June, 2000. [16] S. Imai, "Cepstral analysis synthesis on the mel frequency scale," in proc. of ICASSP'83, IEEE International Conference, Boston, Massachusetts, USA, April, 1983. [17] H. Zen, T. Nose, J. Yamagishi, S. Sako, T. Masuko, A. Black and K. Tokuda, "The HMM-based speech synthesis
system (HTS) version 2.0," in proc. of Sixth ISCA Workshop on Speech Synthesis, Bonn, Germany, August, 2007. [18] S. Young, G. Evermann, M. Gales, T. Hain, D. Kershaw, X. Liu, G. Moore, J. Odell, D. Ollason, D. Povey and others, "The Hidden Markov Model Toolkit (HTK) version 3.4," Cambridge University Engineering Department, December, 2006. [19] P. Boersma and D. Weenink, "Downloading Praat for Windows," 10 September 2013. [Online]. Available: http://www.fon.hum.uva.nl/praat/download_win.html. [20] G. van Rossum and others, "Python language website," World Wide Web: http://www.python.org, 2007. [21] S. Hussain, "Phonological Processing for Urdu Text to Speech System," in Contemporary Issues in Nepalese Linguistics (eds. Yadava, Bhattarai, Lohani, Prasain and Parajuli), Linguistics Society of Nepal, Nepal, 2005. [22] A. Raza, S. Hussain, H. Sarfraz, I. Ullah and Z. Sarfraz, "An ASR System for Spontaneous Urdu Speech," in proc. of Oriental COCOSDA, Kathmandu, Nepal, November, 2010. [23] H. Zen, "An example of context-dependent label format for HMM-based speech synthesis in English," The HTS CMUARCTIC demo, July, 2011. [24] S. Imai, T. Kobayashi, K. Tokuda, T. Masuko, K. Koishida, S. Sako and H. Zen, "Speech signal processing toolkit (SPTK)," 2009. [25] ActiveState and ActiveTcl-User-Guide, "Incr Tk", ActiveTcl 8.4.1.0, Nov, 2002. [26] S. J. Young, J. J. Odell and P. C. Woodland, "Tree-based state tying for high accuracy acoustic modelling," in proc. of the workshop on Human Language Technology, Plainsboro, New Jerey, USA, March, 1994. [27] K. Shinoda and T. Watanabe, "Acoustic Modeling Based on the MDL Principle for speech recognition," in proc. of EuroSpeech-97, Rhodes, Greece, September, 1997. [28] W. D. Voiers, "Diagnostic evaluation of speech intelligibility," Benchmark papers in acoustics, vol. 11, Stroudsburg, Pennsylvania, 1977, pp. 374-387. [29] A. S. House, C. E. Williams, M. H. Hecker and K. D. Kryter, "Articulation-Testing Methods: Consonantal Differentiation with a Closed-Response Set," The Journal of the Acoustical Society of America, vol. 37, no. 1, January 1965, pp. 158-166. [30] U. Jekosch, "The cluster-based rhyme test: A segmental synthesis test for open vocabulary," in Speech Input/Output Assessment and Speech Databases, Noordwijkerhout, 1989. [31] B. Bozkurt, O. Ozturk and T. Dutoit, "Text design for TTS speech corpus building using a modified greedy selection," in INTERSPEECH, 2003. [32] "CMU Sphinx - Speech Recognition Toolkit," Carnegie Mellon University, [Online]. Available: http://cmusphinx.sourceforge.net/html/cmusphinx.php. [Accessed 3 March 2014].